From chemistry-request@server.ccl.net Mon Dec 6 08:04:41 1999 Received: from rzusuntk.unizh.ch (rzumail2.unizh.ch [130.60.128.10]) by server.ccl.net (8.8.7/8.8.7) with ESMTP id IAA17330 for ; Mon, 6 Dec 1999 08:04:38 -0500 Received: [from rzusgi.unizh.ch (chburger@rzusgif.unizh.ch [130.60.68.19]) by rzusuntk.unizh.ch (8.8.5/SMI-5.31) with ESMTP id MAA08236 for ; Mon, 6 Dec 1999 12:56:17 +0100 (MET)] From: "Dr. Peter Burger" Received: [(chburger@localhost) by rzusgi.unizh.ch (8.8.8/SMI-IRIX6.2/USAR_MAIL_V2) id MAA13781 for CHEMISTRY@ccl.net; Mon, 6 Dec 1999 12:56:16 +0100 (MET)] Message-Id: <199912061156.MAA13781@rzusgi.unizh.ch> Subject: Re: CCL:Comparative Linux quantum chemistry benchmark To: CHEMISTRY@ccl.net Date: Mon, 6 Dec 1999 12:56:16 +0100 (MET) X-Mailer: ELM [version 2.4 PL25 PGP2] Content-Type: text Dear Marcel, thanks for your E-mail and comments. > a benchmark usually means checking DIFFERENT programs on ONE machine, or > ONE program on DIFFERENT machines. This makes the benchmark meaningful. I agree, however, due to license restrictions, life is not always that simple in that I could not perform all the benchmarks on an identical machine. However, where it was possible, the benchmarks data displayed an excellent correlation with the clock frequency (slope of .99) in that that seems not to be too much of a concern if at all. > BTW, are the default grids the same for all programs ? Do they all use > numerical integration, and if so, are the energies all the same, up to the > same accuracy ? Is the numerical integration valid up to the same accuracy, > etc. etc. ? Energies are hard to compare since different basis sets _have_ to be used STO's for ADF, GTOs for most of them and numerical for Dmol.. The default grids were not always identical, some of them use modified grids, which cannot always be controlled by the user.. Essentially, it's better to see this comparison more as an applied benchmark for applied chemists using the default settings. Even more matters actually the number of cycles that are required with the same convergence criteria. This may vary from 10-28 cycles!!! and that the programs achieve convergence at all! which is definitely _not_ the case for transition metal complexes I am interested in. > > Else, these times do not say anything more, than that you have had a nice > > time, the afternoon you were playing with these programs. For me, it's actually more than _playing_ since time for completion matters to me. However, obviously there are more points to that including the quality of optimizers, SCF-convergence, flexibility of the programs, for instance to do an unrestricted/Cosmo/ECP run on open shell transition metal systems is not all that simple for a number of programs which I had to learn the hard way. So, the message is get the job done! For instance the ridft method in Turbomole is so fast in parallel runs (and still accurate) that you can treat even large systems fully quantum mechanically and have not resort to hybrid methods. Peter ------------------------------------------------------------- Peter Burger Anorg.-chem. Institut Universitaet Zuerich chburger@aci.unizh.ch